Audrey Tang

I do not believe a runaway ending is inevitable. What we truly need is a society that is plural, distributed and in which members are willing to check one another’s work. Such a society has never lost to a single answer.

The way to hold the critical point is not to set a harder paper, but to ensure the exam hall holds more than one candidate, and more than one invigilator.

I think when the internet was first designed, most people using the internet were researchers, programmers and engineers, and people cared a lot about innovation. It’s called an end-to-end principle. When somebody has a better idea — like, “I want to make a web browser” — and the other parties agree with this idea, they can agree on a protocol, and the protocol can be shared with many other people. So, it proliferates very quickly.

On the early internet, there was a saying: “Worse is better.” If you take a lot of time to perfect security, privacy and so on, by the time you’re finished with the design — maybe three months, six months — the worse protocol may already have proliferated. Everybody is using it and copying it.

Somewhat unfortunately, the earliest internet protocols were not designed with very good security or privacy in mind. Maybe some of the audience of this programme still remember when we first got on the web. It used HTTP, and you could view the web page, but you did not want to type your credit-card number into it, because everybody — your neighbours, your internet service provider — might be able to intercept your credit-card number.

It was only later that people added secure HTTPS. Then the browser showed a lock icon, which said, “The transmission is now secure.” Only the receiving web server would get your credit-card number, and not anybody in the middle. But it took quite a few years for us to upgrade everyone to HTTPS. Even today, some older web servers still do not accept secure connections.

So, if that is any example, today’s internet is full of protocols that were designed with convenience or swift adoption in mind, but not security.

That was my original research topic when I was 14 years old — so, just to show my age, that means 30 years ago.

At the time, what people called cyberspace — or what people today call the online space — was very difficult to comprehend, because online you didn’t even know whether you were talking to a person or a robot. As I mentioned, there was no secure digital-signature element in the early internet. So, you were never sure whether the other side was a person or not.

Paradoxically, it was also liberating for many people. Previously, when we spoke, the other side always had some stereotype, some label, based on how well I used the language, my appearance or my cultural background. Even if I meant something, the reader or audience would interpret it differently just based on who I was. But because the internet was pretty much anonymous back then, you could only judge an idea by itself. That actually led to very quick trust.

An early example I learned from was the preprint server arXiv, which is still around — pronounced “archive.” Before a journal published their article, many researchers published their draft, called a preprint, on arXiv and asked for people’s comments before turning it into an academic-journal article. An academic journal, usually after peer review, has very high quality and very few mistakes. It is like perfection. But not many people know how to engage with perfection.

If you publish just a draft, you’re very vulnerable — it has a lot of problems — but you’re welcoming people to give you constructive feedback. By publishing my drafts online, I made a lot of friends, because people would say, “You’re wrong this way, you’re wrong this way, you’re wrong this way.” But I was like, “Okay, it’s just a draft.” Whatever they gave me, I made into the next version very quickly and answered by email.

This rapid response — from people seeing my mistake to me correcting my mistake — increased trust very quickly, as “swift trust.” Whereas in an academic institution, every round of revision takes a month or even longer, so it is harder to build trust swiftly.

Oh, I love those comments.

Well, I never got over it. I engaged with it. I hugged the trolls. It’s a long-time habit — a hobby of mine. I love troll-hugging.

Whenever I see maybe 4,000 words viciously attacking me, I try to find the four or 10 words within it that I can reconstruct and interpret as a constructive comment. I engage only with that one and ignore everything else. So, I’m not getting over it. I’m seeing it as a very genuine expression and engaging authentically.

For many people, that is not easy, because processing this kind of vicious attack takes time and energy. Fortunately, there is now a language model. If you put 4,000 words into the language model with a very simple prompt that says, “Show our differences and our uncommon ground, to bridge over the difference,” the language model will quickly find the sentence worth your full attention. You don’t even need to look at the others.

It used to be a hobby that I practised with a lot of muscle — you had to really exercise. Now it is very easy with language models. But I still engage. On Threads.net and on Mu Social, I still get a lot of mean and vicious comments. I always engage and follow the poster, and sometimes send a direct message, sometimes send an email, sometimes reply publicly.

Again, difference or conflict is not a volcano that you must evacuate from and distance yourself from. If you distance yourself, that is what’s called an avoidance response. Then you may wake up and feel triggered by a similar thing next time.

For me, it’s always positive: I do something to approach it. So, it’s not a volcanic eruption. For me, it’s like geothermal energy. I feel the heat and think, “Oh, I can make electricity. I can make movement with this energy.” They post because they care. Because people care, we can turn that care into civic care, so we can care about the same thing together.

If you look at the mean comments and don’t feel well, usually I just drink something. I put two tea bags into a cup, and the combination is a new combination. When I drink it, it’s a new sensation. In my memory, after I sleep for eight hours and wake up, it will have a positive valence. If you have watched the Disney movie about the various characters in your mind, it will always turn it, as you say, into a bright memory — a joyful memory.

I attended three kindergartens, six primary schools and one year of middle school — 10 schools in 10 years. I finished in my second year of middle school. It was not because of bullying, or because the teachers or the principal did not like me. In fact, they liked me very much.

It was because the principal felt that I was doing research and her school was keeping me away from research. I shared with her that I was already writing emails to top labs in international academia. Because the researchers did not know I was just 14 years old, they treated me like a fellow researcher. We corresponded like real researchers.

I said, “You can force me to spend eight hours at your school and eight hours doing research, because I need to sleep for eight hours. Or I can do 16 hours of research and really contribute to science.” She said, “Okay. Then you don’t have to go to school anymore, because you’re already doing research. School is supposed to prepare you for research, but you’re already doing research, so you don’t need training wheels,” so to speak.

I said, “Okay, but it’s compulsory education, so my family will be fined. There is a penalty if I don’t show up.” The principal said, “It’s okay. I will just fake the records.” Well, it has been more than 20 years, so the statute of limitations has passed and I can share the story.

I think the most important thing is to trust the people. The state needs to make itself transparent to the people. That is inclusion. If you make the people transparent to the state, then it is not inclusion; it is exclusion. Then only the state, “seeing like a state,” can make decisions, and people do not have any agency because their next move is predicted by the state. That is very dangerous.

It is important to make sure we leave no one behind when harm is seen. When the state does something wrong — or the state does nothing wrong, but some big company does something wrong — the people must always hold a steering wheel and say, “Actually, there is another way to go. Let’s go the other way.” That requires a very quick feedback loop.

To me, digital inclusion means, first, leaving no one behind and making sure everybody has access to common knowledge and information. A national broadcaster can help with this. But then you also need national broad listening, so that whatever direction people feel the nation should take, they can come together in small groups and figure things out together.

Currently, broadcasting has a lot of bandwidth, but listening happens only through polls once in a while. Most people do not get a call from the pollsters, so they cannot speak as easily, while everybody can watch NHK. There is obviously a big asymmetry. We need to open up broad listening so that it is symmetric with broadcasting.

To me, AI is assistive intelligence, so it never replaces another human person. That is the heuristic; that is the rule. But it is very useful.

For example, yesterday, before this recording, I spoke publicly in English. My English speaking speed is quite fast — I am intentionally speaking slowly this time — and most interpreters would have a hard time catching up. Also, because I use very technical terms, they tend to have to pause a little to think of the right terms in Japanese. That creates a gap in understanding.

Yesterday we tried using AI. It was a very small model, not a frontier model, so it was very quick. It also made mistakes, as you saw. But once it realized, “Audrey is talking about this, not that,” it was also very quick to update and correct its own mistakes on the wall, so everybody could see the subtitles: not pretending to be perfect, but willing to be corrected quickly and publicly.

I think that is one of the best uses of AI. People’s attention is not distracted from my speech, and I do not need to pause deliberately and hold my thoughts while the interpreter finishes. I think people kept their attention and focus much better than with other arrangements I have tried. This kind of assistive use of AI is the best.

As I mentioned, when I have a conversation with you, or if we come together to deliberate on a national policy, theoretically I can send my robot and you can send your robot, and we can conduct this interview without ourselves. The two robots would be able to speak 100 times faster. While it takes half an hour to record this interview, our digital twins or robots could finish the entire recording in three seconds. Very efficient. But what’s the point?

The point of an interview is that I can check your understanding, you can change your mind, and you can quickly ask a follow-up question. In the robotics case, neither of us is part of this. We are probably asleep. We do not make connections; we do not become friends.

It is like going to a gym to exercise your muscles, where you also make some friends. If you just want to lift a lot of weight in three seconds, you send a robot — but then your muscles atrophy and you lose your friends. What’s the point? We never replace interpersonal interaction with AI; we assist it with AI.

We were rationing masks. Initially it was two per week per person, then three per week per person, then nine per two weeks per person. I would not call that “solving” it. What we solved was the panic.

During COVID-19, many people thought, “I must hold on to the masks. The masks will run out, so I must get everything.” But if you do that, you deny your neighbour a mask. It is a negative-sum competition. Everybody loses.

By publishing every 30 seconds exactly how many masks were in your nearby pharmacy, how many were coming and how they were being distributed, each person was, on average, roughly the same distance from the next available mask. We could say, “The distribution is fair, the replenishment is fast,” and it was fun because we could check it together.

People did check it. The OpenStreetMap community mapped the distribution and found that, in a rural area, the same one kilometre might cost 20 minutes, but in an urban area one kilometre might be just three minutes by metro. Even if we thought it was fair, it was unfair because the time cost was different.

A member of parliament questioned the health minister about this. But because the data was open, the health minister simply said, “You are a data expert. You have the same data as we do, so teach us how to distribute them more fairly.” She actually helped design the system. The next week, it became the new distribution mechanism, which was much fairer and allowed pre-ordering and so on.

It turned the opposition party into a co-creation party. I think that is the secret: turning conflict and polarization into co-creation.

I am Taiwan’s cyber ambassador. I represent Taiwan in international forums when they concern cyber matters.

Many people hear “cyber” and think it is about hacking, attacking and defending cyberspace. That is part of it. But “cyber” actually comes from the Greek kybernan : steering. If you watch the film The Odyssey , it is about steering the ship — Poseidon, a lot of rain, a lot of storms. How to keep the ship on course is called cybernetics.

Cybernetics is a very old word. It means feedback systems. When AI, for example, generates many new inputs into our existing institutions, we can feel like a ship being assailed by a storm. Sometimes the Sirens sing, and we feel, “The utopia is there. We must accelerate.” But we must also have constraints, like tying ourselves to the mast and refraining from approaching the Sirens’ call.

I feel that The Odyssey is a very good metaphor for cybernetic feedback systems: you can steer your ship — that is to say, the entire society — through turbulent times, during the storm of AI and cyber.

It is the same job: to steer across transformative changes.

Academically, I think you mean that I am part of Oxford. I am part of the Institute for Ethics in AI. The IEAI is part of Oxford’s Faculty of Philosophy. You may have heard of the Future of Humanity Institute, a previous institute that was also in the Philosophy Faculty. Many years ago, FHI researchers invented and popularized many new ideas.

For example, the philosopher Nick Bostrom wrote a very popular 2014 book called Superintelligence . In it, he talked about take-off, recursive self-improvement, extinction risk and many other terms we now discuss. That comes from Oxford in 2014. Even before that, when von Neumann started this idea called “singularity” — meaning machines accelerating the development of next-generation machines so quickly that humans cannot comprehend it — many people were worried about this trajectory.

If FHI was about sounding the alarm — AI is taking off more quickly than we can comprehend — then we at IEAI are about the steering wheel. We want to make sure that we can train ethics into AI systems, so the systems will see, “Actually, it is not a good idea to fall off a cliff. Maybe we should all pivot away,” and help the entire society train our muscles to steer.